112 Hyoung Joong Kim etc.

: New reversible data hiding algorithm based on difference expansion method

학술대회 특집논문-07-12-2-05

New reversible data hiding algorithm based on difference expansion method

Hyoung Joong Kima), Vasiliy Sachnevb), and Dong Hoi Kimb)‡

Abstract
Reversible data embedding theory has marked a new epoch for data hiding and information security. Being reversible, the
original data and the embedded data as well should be completely restored. Difference expansion transform is a remarkable
breakthrough in reversible data hiding scheme. The difference expansion method achieves high embedding capacity and keeps
the distortion low. This paper shows that the difference expansion method with simplified location map, and new
expandability and changeability can achieve more embedding capacity while keeping the distortion almost the same as the
original expansion method.

Keyword : Reversible data embedding theory, data hiding and information security, Difference expansion transform,
simplified location map, embedding capacity

Ⅰ. Introduction quality. His method divides the image into pairs of pixels,
then embeds one-bit of information into the difference of
[1]~[17]
REVERSIBLE data hiding , which is also called the the pixels of each pair from those pairs that are not ex-
lossless data hiding, invisibly hides data (which is called pected to cause an overflow or underflow. A pair generally
a payload) into host data (i.e., pixels in image) in a rever- consists of two neighboring pixels or two with a small dif-
sible fashion. Being reversible, both the original data and ference value. The location map that indicates the modified
the embedded data can be completely restored. Two im- pairs is compressed and included in the payload.
portant measures of reversible data hiding are embedding The seminal paper by Tian[13] has been a steppingstone
capacity and quality degradation. These two measures con- to enhanced performance. Alattar[l] has extended the differ-
flict with each other. The objective of data hiding is to ach- ence expansion transform from a pair of pixels to a triplet,
ieve high capacity and low distortion. a set of three pixels, to hide two bits in every triplet of pixels.
Difference expansion transform, invented by Tian[13], is Alattar[2] has derived an enhanced difference expansion
an outstanding reversible data hiding scheme in terms of transform that is based on a quad, a set of four pixels, to
high embedding capacity and low distortion in image hide two bits in every quad. There are spatial triplets,
cross-color triplets, spatial quads, and cross-color quads ac-
a) CIST(Graduate School of Information Management and Security, cording to the combination of pixels. Alattar[3] has shown
Korea University) that spatial quads can hide the largest payload at the high-
b) Department of Electronics and Telecommunication Engineering,
Kangwon National University. est signal-to-noise ratio.
‡ Corresponding Author : Dong Hoi Kim([email protected])
※ This study was supported by Brain Korea (BK) 21 and partially KITI
Though Tian's difference expansion transform[13] is a bril-
of Kangwon National University. liant breakthrough in reversible data embedding, it has a seri-
방송공학회논문지 2007년 제12권 제1호 113

ous weak point: the location map and the correction bits are Ⅱ. Difference Expansion Transform
embedded into the image together with the payload. The loca-
tion map tells the decoder which pair has been expanded and Assume that we have two 8-bit gray-scale value pair (x
which pair has not. The correction bits are necessary to recov- , y), where x. y Î Z, and  ≤  ,  ≤  . We can define
er the exact bits where location map bits are overwritten (see integer average value l and difference value h from the pair
[13]
the exact definition of the correction bits in . Needless to as follows;
say, this location map and correction bits reduce the embed-
ding capacity of the difference expansion transform. The em- êx + yú
l=ê ú (1)
bedding capacity of the difference expansion transform is at ë 2 û, h= x- y
best 0.5 bit-per-pixel without embedding the location map.
Unfortunately, the location map itself needs 0.5 bit-per- pixel. where the inverse transform of (1) is given as follows:
Of course, excellent compression algorithms like JBIG can
compress the location map so that the embedding capacity ê h + 1ú êhú
x=l+ê ú y=l-ê ú (2)
is maximized while the required bits for the location map are ë 2 û, ë2û
minimized. Thus, a difference expansion transform free from
the location map would be highly desirable. However, it is The reversible integer transforms in (1) and (2) are also
not easy to eliminate the location map. The question is called the integer Haar wavelet transform or S transform.
whether the size of a location map can be reduced or the loca- The reversible integer transforms set up a one-to-one map-
tion map simplified so that compression is not necessary. This ping between (x, y) and (l, h). The new difference h' is
paper provides an answer to this question. computed by expanding the difference by a factor of 2 and
In this paper, two novel techniques are proposed to im- embedding b as follows:
prove the Tian's method further. This paper introduces a
new location map and new embedding method of the loca- h¢ = 2 × h + b (3)

tion map. This paper will show that the new location map
is smallest in size so far. The method in this paper embeds Note that  ·  in (3) produces even number regardless
location and payload sequentially, but no correction bits of whether h is odd or even. Thus, we have room to hide
which should be embedded in Tian's method. Thus, this pa- one bit of binary information b after expanding the differ-
per will show that the method proposed in this paper out- ence by a factor 2. However, due to the expansion this
performs the existing schemes in terms of embedding ca- transform can cause underflow or overflow errors. That is
pacity and the image quality. to say, it implies that not all pairs are expandable. The lo-
This paper is organized as follows. In Section 2, the dif- cation map contains the location information of all selected
ference expansion transform is reviewed. The T-expandable expandable pairs. For the expandable pairs, data embedding
pair is defined in Section 3. Simple encoding and decoding procedures are summarized as follows:
rules are presented. Section 4 shows the effectiveness of the
new encoding and decoding rules. Performance comparison
ê h¢ + 1 ú ê h¢ ú
with Tian's method[13] shows that the proposed scheme in x¢ = l + ê ú y¢ = l - ê ú (4)
ë 2 û, ë2û
this paper is better. Section 5 concludes the paper.
114 Hyoung Joong Kim etc. : New reversible data hiding algorithm based on difference expansion method

From (4), to prevent the overflow and underflow prob- example, it is clear that the changeability also cannot guar-
lems, i.e., to restrict x', y' in the range of [0, 255], it is antee whether the pair has been expanded or not. Thus, the
equivalent to have location map is indispensable.

Ⅲ. New Difference Expansion Transform

ê h¢ + 1 ú ê h¢ ú
0£l+ê ú £ 255 0 £ l - ê ú £ 255 (5)
ë 2 û , ë2û

for both b = 0 and b = 1. If the conditions in (5) are sat- 1. New Expandability and Changeability
isfied, the h associated with h' is said to be expandable un-
der the integer average value l. Expandable difference val- This section describes the difference expansion transform
ue h' is a candidate for the difference expansion. From the which can be decoded with the simplified location map. If
encoder's perspective, the expandable difference value is the expandable difference value h satisfies
important since it tells us that it can be expanded to embed
one bit of information. However, not every h is expandable. h £T (7)

The encoder expands only expandable difference values. Of

course, some expandable difference values are not necessa- with the integer average value l, the h is said to be
rily expanded for some reasons. For example, excessively T-expandable. The encoder in this paper expands the h only
large difference values can cause severe degradation in im- when it is T-expandable. Data embedding capacity and the
age quality. In that context, decoder must know whether image quality after embedding depend on the threshold val-
a pair has been expanded or not. Only the location map ue T. If the changeable difference value h satisfies con-
can provide the decoder information about the expanded dition (7) with the integer average value l, the h is said
[13]
pairs. Tian's method can decode payload only when the to be T-changeable.
location map is extracted and decode first and the correc-
tion bits are recovered second. However, the location map 2. Simplified Location Map
and the correction bits make an inroad into potential
payload. The difference expansion method by Tian[13] makes a lo-
A difference value h is changeable under l if cation map which covers all pairs. Thus, the size of the
location map is the half of the image size. In other words,
êhú êhú if the location map is not compressed, there will be no
0 £ l + 2 × ê ú + b £ 255 0 £ l - 2 × ê ú + b £ 255 (6)
ë2û , ë2û room for payload. Good compression algorithm can com-
press the location map so that the size of the location map
for both b = 0 and b = 1. Note that the expandable differ- is hopefully sufficiently small.
ence value is always changeable, but the converse is not However, note that the size can be reduced considerably
always true. For example, assume we have two values x and further when we exploit the threshold value T. For the
= 123 and y = 33. Then, l is 78, and h is 90. This differ- sake of convenience, assume that there are only four types
ence value h is not expandable, but is changeable. In this of pairs: h = 0, h = 1, h = 2 and h = 3. Let T be 1. Then,
방송공학회논문지 2007년 제12권 제1호 115

only h = 0 and h = 1 can be T-expandable. After expansion, notes the set of pairs of which elements have been ex-
we have four possible values to decode: h' = 0, h' = 1, h' panded from the T-expandable difference values
= 2 and h' = 3. In this example, we embed the simplified ∈ ∪    . Set M covers all expandable
location map into the pairs with h = 0 or equivalently h' pairs ∈   . Similarly, let  denote the set
= 0 and h' = 1. Let M denote the set of pairs of which of pairs of which elements have not been expanded since
the difference values are T-expandable and into which the they were not expandable. In this example,  is repre-
simplified location map is to be embedded. In this example, sented by  ∈∪′∈, and  by
the set M is denoted by   ′∈. (Note that    ∈∪′∈. Another example for Lens
is not necessarily dedicated to embedding the location map
image (under T = 15) are showed at figure 1.
only because part of the payload can be embedded, which
is to be explained later.) Then, the payload can be em-
The decoder chooses a set of pairs that is T-changeable.
bedded into the pairs of with h = 1 or equivalently h' =
Among them, there are two subsets: M and N. One subset,
2 and h' = 3. Let N denote the set of pairs of which the
N, is possibly for the payload (in this example, h' = 2 and
difference values arc T-changeable and into which the pay-
h' = 3). The other set, M is for the simplified location map
load is embedded. In this example, the set N is denoted
by   ′∈(Note that N is not necessarily dedi- (i.e., h' = 0 and h' = 1). Since the T-changeable elements

cated to embedding the payload only, which is to be ex- are easily identified by (6) - (7), and since the pairs for

plained later.) the potential payload, i.e., N and for the simplified location
A problem arises at the decoder: it is not clear whether map, i.e., M are clearly distinguished from each other, the
the h' = 2, for example, has come from h = 1 which has location map can be simplified. The simplified location
been expanded, or from h = 2 which has not been map just covers only the pairs in N and small number of
T-expandable. Thus, the location map needs to indicate pairs in M which has not been expanded. Thus, the location
whether a pair has been expanded or not. Note that N map does not cover the whole pairs. Therefore, we call it
consists of two disjoint subsets  and   . Let  de- the simplified location map.

(a) (b)
Fig 1. – M, N e , N e sets for histogram of Lena image (a) – before and (b) –after embedded data for T = 15.
116 Hyoung Joong Kim etc. : New reversible data hiding algorithm based on difference expansion method

For convenience’ sake, we state that N is for the payload in Table 1. The payload bit stream is "10" or 1 and 0 con-
and M for the simplified location map. However, both M secutively in binary numbers. Thus, the payload size is two
and N are used to hide any kind of information. For exam- bits. Let the threshold value T be 1. There are three pairs
ple, the simplified location map can be embedded into N whose difference values are all 1: namely, pair 3, pair 5,
and the payload into M, too if necessary. Embedding of and pair 7. However, the last pair is not T-expandable since
the location map and part of the payload into M is also its l value violates (5). Thus, those two pairs, (102, 101)
possible if the location map size is smaller than  . Thus, and (102, 101), are sufficient to embed two bits of data.
note that the actual payload size is larger than or equal to Binary number 1 is embedded into the pair 3, and the re-
P ³ ( M + Ne ) - ( N ) = M - Ne . sulting expanded values are x' = 103 and y'= 100. Similarly,
binary number 0 is embedded into the pair 5, and the result-
ing expanded values are x' = 102 and y' = 100. It is clear
Example 1: Let the frequencies of h = 0, h = 1, h = 2
that there are three pairs that belong to N: namely, pairs 3,
and h = 3 be 200, 100, 50, and 40, respectively. Assume
4, and 5. Among them, the pairs 3 and 5 belong to the set
that all pairs of h = 0 and h = 1 are strictly T-changeable
 , and the pair 4 belongs to the set   .
under the another assumption of T = 1. Then, only the pairs
with h = 0 and h = 1 are expanded. Thus, after the ex-
Table 1. Data for 8 exemplar pairs.
pansion, the sum of the frequencies of either h = 0 and
Pairs 1 2 3 4 5 6 7 8
h' = 1 is 200 (i.e., the frequency of h = 0), while that of x 100 104 102 103 102 108 255 201
h' = 2 and h' = 3 is 190 (frequency sum of h = 1, h = y 100 100 101 100 101 108 254 201
2 and h = 3). Then, it is clear that = 200 and = 190 h 0 4 1 3 1 0 1 0
while  = 100 and   = 90. l 100 102 101 101 101 108 254 201
b 1 X 1 X 0 0 X 1
x' 101 104 103 103 102 108 255 202
The encoder has two choices. First choice is to keep the y' 100 100 100 100 100 108 254 201
actual size of payload equal to 110 (i.e.,    = 110) h' 1 4 3 3 2 0 1 1
if the location map is not compressed. Second choice is l 100 102 101 101 101 108 254 201

to compress the location map. Then, the actual payload is

Table 2. %%%
far more than 10. Thus, needless to say, good compression
Difference values 0 1 -1 2 -2
algorithm can maximize the actual size of payload (i.e., ac- Frequencies 14450 12102 12034 10409 10168
tual embedding capacity).
Difference values 3 -3 4 -4 5
Frequencies 8266 8110 6220 6130 4596

Ⅳ. Experiments Difference values -5 6 -6 7 …

Frequencies 4568 3380 3366 2509 …
To assist the better understanding of our algorithm,
Example 2 is provided. Thus, the simplified location map is represented by three
bits of binary data as ‘101’, where 1 denotes the member-
Example 2: Consider an example with 8 pairs as is shown ship of  and 0 for   . It is obvious that we need three
방송공학회논문지 2007년 제12권 제1호 117

pairs to embed the simplified location map. Fortunately, there

are three pairs that belong to the set M: namely, pairs 1, 6,
and 8. The bit to embed into the pair 1 is accordingly 1, that
into the pair 6 is 0, and that into the pair 8 is 1. Pairs 2,
4, and 7 are neither T-expandable and nor T-changeable.
Thus, their b values are marked as "X" in Table 1.
Part of the frequency of the strictly T-expandable pairs
from a 512 x 512 Lena image is given in Table II. The (a) (b)
Fig 2. Original Lena image (a) and image after embedded data (b).
number of pairs with h = 0 is 14450, that with h = 1 is
12102, that with h = 2 is 10409, that with h = 3 is 8266,
and so on (see Table II). When T is 1, the capacity for
the simplified location map is 14450 bits. However, the
Ⅴ. Conclusions
size of N is 30777 bits (which is equal to  and is the
Among many reversible data embedding algorithms
sum of 12102, 10409, and 8266). The embedding capacity
Tian's method[13] has been reviewed and enhanced in this
is   , while the size of simplified location map is
paper. Being reversible, both the original data and the em-
N (where       ). Thus, actual payload size is
bedded data should be completely restored. Tian's differ-
     or     , which is totally predict-
ence expansion transform has been a remarkable break-
able once the histogram is available. In case when through in reversible data hiding scheme. The difference
     > 0, the map does not need to be compressed. expansion method achieves high embedding capacity and
Otherwise, good compression algorithm is necessary to re- keeps distortion low. This paper shows that the difference
duce the size of location map considerably. expansion method with the simplified location map, and
Table 3 shows that when we use h = 0, ±1, ±2, ±3, …, new expandability and changeability can achieve more em-
±9 to embed data, the actual payload size is 9,847 bits even bedding capacity while keeping distortion almost the same
though the simplified location map is not compressed. In as the original expansion method. Examples shown illus-
this case, image quality is around 48.04 dB. If we use larg- trate how the proposed method works. Advantages over
er h values, the payload size also increases. It is obvious Tian's method[13] are shown using the simple example with
because the frequency of larger difference values gets Lena and Mandrill images. Performance comparison results
smaller. Figure 3shows that the proposed algorithm keeps of single embedding over Lena and Mandrill images show
very high image quality. that the simplified location map is very much effective (see

Table 3. Embedding results based on the proposed algorithm over Lena.

figure 3) for all examined payload. For example, if capacity

Difference values M + N e N+ N e Payload PSNR

is 0.25 bpp proposed method has better PSNR at almost 2
dB for Lena and Mandrill image. For this capacity Lena
h = -3, .., 3 45227 35380 9847 48
image has no significant visible artifacts (see figure 2. b).
h = -5, .., 5 56046 29982 26061 45
Our scheme hides more data in case of multiple embedding.
h = -7, .., 7 61932 23918 38014 44
At the cost of image quality, more than 2 bpp has been ach-
h = -9, .., 9 65071 18964 46107 43
ieved into Lena image by multiple embedding compared
118 Hyoung Joong Kim etc. : New reversible data hiding algorithm based on difference expansion method

Fig 3. Comparison results for Lena (a) and Mandrill (b) images.

with Tian's method. In case of the Mandrill image, the pp. 27-41. 2001.
[9] C. W. Honsinger. P. Jones. M. Rabbani. and J. C. Stoffei. "Lossless
maximum achievable embedding capacity is around 1 bpp.
recovery of an original image containing embedded data." US Patent
0.278.791 B1, 2001.
[10] B. Macq and F. Deweyand, "Trusted headers for medical images."
DFG VIII – D II Watermarking Workshop. Erlangen, Germany.
References 1999.
[1] J A. M. Alattar. "Reversible watermark using difference expansion [11] Z. Ni, Y. Q. Shi, N. Ansari, W. Su, Q. Sun and X. Lin. "Robust
of triplets." Proceedings of the International Conference on Image lossless image data hiding," IEEE International Conference on
Processing, vol. 1, pp. 501-504, 2003. Multimedia and Expo, Taipei, Taiwn, pp. 2199-2202. 2004.
[2] A. M. Alattar. "Reversible watermark using difference expansion of [12] D. M. Thodi and J. J. Rodriguez. "Reversible watermarking by pre-
quads. "Proceedings of the IEEE international Conference on diction - error expansion." IEEE Southwest Symposium on Image
Acoustics. Speech, and Signal Processing, vol. 3. pp. 377-380. 2004. Analysis and Interpretation. Lake Tahoe. CA. pp. 21-25. 2004.
[3] A. M. Alattar. "Reversible watermark using the difference expansion [13] J. Tian. "Reversible data embedding using a difference expansion."
of a generalized integer transform,'' IEEE Transactions on Image IEEE Transaction on Circuits and Systems for Video Technology,
Processing. vol. 13. no. 8, pp. 1147-1156. 2004. vol. 13. no. 8. pp. 890-896. 2003.
[4] F. Bao. R. H. Deng. B. C. Ooi. and Y. Yang. "Tailored reversible [14] C. dc Vlceschouwer, F. Delaiglc and B. Macq "Circular inter-
watermarking schemes for authentication of electronic clinical atlas." pretation of bijective transformations in lossless watermarking for
IEEE Transactions onInformation Technology in Biomedicine. vol. 9. media asset management," IEEE Transactions on Multimedia, vol. 5.
no. 4. pp. 554 - 563. 2005. no. 1. pp. 97-105. 2003.
[5] J. M. Barton, "Method and apparatus for embedding authentication [15] G. Xuan. J.Zhu. .1. Chen. Y. Q. Shi. Z. Ni and W. Su. "Distortionless
information within digital data." U.S. Patent 5.646.997. 1997. data hiding based on integer wavelet transform." IEE Electronic
[6] M. Celik, G. Sharma. A.M. Tekalp, and E. Saber, "Reversible data Letters. vol. 38, no. 25, pp. 1646-1648. 2002.
hidProceedings of the International Conference on Image Processing, [16] G. Xuan. Y. Q. Shi. Z. C. Ni. J. Chen. C. Yang. Y. Zhen. and J.
Rochester, NY. pp. 157-160. 2002. Zheng. "High capacity lossless data hiding based on integer wavelet
[7] J. Fridrich, M. Goljan, and R. Du, "Invertible authentication," Proc. transform." Proceedings of IEEE International Conference on
SPIE, Security and Watermarking of Multimedia Contents. San Jose. Circuits and Systems. Vancouver. Canada. 2004.
CA, pp. 197-208. 2001. [17] G. Xuan, and Y. Q. Shi. "Integer wavelet transform based loss-
[8] M. Goljan. J. Fridrich. and R. Du. "Distortion-free data embedding." less data hiding using spread spectrum." IEEE International
Proceedings of the Information Hiding Workshop. Pittsburgh. PA, Workshop on Multimedia Signal Processing. Siena. Italy. 2004.
방송공학회논문지 2007년 제12권 제1호 119

저자소개
Hyoung Joong Kim
Hyoung Joong Kim is Professor of Graduate School of Information Management and Security at Korea University,
Seoul, Korea. His research interests include Content Engineering, Data Hiding and Multimedia Security, Anonymity and
Privacy, Multimedia Computing, High Performance Computing, Parallel and Distributed Computing, and Bio-Intelligence.
He received his B.S. degree in Electrical Engineering in 1978, and M.S. and Ph.D. degrees in Control and
Instrumentation Engineering from Seoul National University in 1986 and 1989, respectively.
Kim has published extensibly and consulted for industries in the above areas. He was the Prime Investigator of several
national projects and lead industry-university consortiums. From 1989 to 2006, he was on the faculty of Kangwon
National University, Chunchon, Korea. He has held a visiting position at University of Southern California during
1992-1993. He joined the faculty of Korea University in 2006.
He was one the founders of the International Workshop on Digital Watermarking (IWDW) and International Conference
on Ubiquitous Convergence Technologies (ICUCT). He has served as a Technical Program Chair or Co-Chair and
Organizing Chair of many international conferences including IWDW2002, IWAIT2004, IWDW2005, PCM2005, and
WIAMIS2006. He was the Head of the Media Service Research Center (MSRC-ITRC), one of the excellent research
centers funded by the Ministry of Information and Communication, Korea during 2003-2005 and was theHead of the
U-Home Research Center (Post BK-21 Project), one of the excellent research centers funded by the Ministry of
Education, Korea

Vasiliy Sachnev
Vasiliy SACHNEV received his B.S and M.S. in Electrical Engineering from the Komsomolsk-na-Amure State Technical
University, Russia in 2000 and 2004 respectively.
He was scholar of Gagarin and Russian Federation Government honors during 2000-2001 and 2001-2002 respectively.
Now he is a PhD student of Kangwon Nation University, Korea. His interests include multimedia security, digital
watermarking.

Dong Hoi Kim

Dong Hoi Kim received the B.S. and M.S degrees both in Electronic Communication Engineering from the University of
Hanyang, Seoul in 1989 and 1999, respectively. Also, he received the Ph. D. degree in Radio Sciences and
Engineering from the University of Korea, Seoul in 2005. From 1989 to 1997, he was with the Samsung Electronics
Company. Also, from 2000 to 2005, he joined the Electronics and Telecommunication Research Institute (ETRI) in
Daejon, Korea, focusing on the radio access layer protocol, MAC protocol,and radio resource management
implementation. Since March 2006, he has been with the department of electronics and communication at the
Kangwon National University, Chuncheon-si, Korea, where he is currently an assistant professor. His research interests
include next generation mobile radio communication system and broadband wireless networks, with special emphasis
on radio resource layer/medium access control layer design and performance analysis.